Benchmarks

Benchmarks

CLPeak

• Source: CLPeak
• Runs single precision, global bandwidth and transfer bandwidth benchmarks

Results:

BabelStream

• Source: BabelStream
• Build with: make -f OpenCL.make
• See also https://github.com/doe300/VC4CL/issues/82

Results: sudo ./ocl-stream --float -n 12 --arraysize 640000

Function    MBytes/sec  Min (sec)   Max         Average
Copy        14.707      0.03481     0.03481     0.03481
Mul         26.906      0.01903     0.01903     0.01903
Add         39.416      0.01948     0.01948     0.01948
Triad       39.135      0.01962     0.01962     0.01962
Dot         79.948      0.00640     0.00640     0.00640

Other benchmarks

• mixbench: default work-group size of 256, can be specified as parameter. Hangs on throwing an CompilationError in pre-compilation (see this ??)

• OpenDwarfs: compilation errors in benchmark code

• PolyBench: tested some benchmarks, partially working, not so meaningful output

• SHOC: Some tests require bigger work-group sizes, other have compilation errors

• ViennaCLBench

• FinanceBench: Only benchmarks Black-Scholes and Monte-Carlo support OpenCL. Huge speedup (too much?? 2000-times!) on GPU, but no/wrong results, Mote-Carlo fails to compile (long)

• http://www.bealto.com/gpu-benchmarks_intro.html: crashes at/after CPU-side memory benchmark (copyN)

• gearshifft

• LuxMark

• Bullet benchmark

• chai

• uCLbench: buffer_bandwidth succeeds, others fail in compilation/loading source code, arith_speed fails for some values (e.g. inf)

• OpenCL STREAM